On Tue, 15 Aug 2000, Mike W Ellwood wrote:
> However, one benefit that we as a site were hoping to get from AFS
> seems not to be quite as good as we had thought: replication of
> volumes.
>
>
> Unless I am missing something, one can only replicate read-only copies
> of volumes. That is quite a restriction.
> Or am I wrong?
This is basically correct.
> Just to give me a flavour of what I might possibly be missing,
> would anyone care to (briefly) summarise their own strategy regarding,
> or use of, AFS volume replication?
As others have noted, while AFS admins can maintain replication of the top
level tree, most people dislike managing replication directly.
To deal with this reality and still get some benefits of replication for
data that should be replicated (mostly web html files and applications),
we have created many volumes that are explicitly mounted readwrite (-rw).
Cron scripts periodically scan these readwrite volumes and auto-release
those that have changed. If the server goes down, we [manually] switch to
a parallel version of the parent directory constructed using default
[readonly] mountpoints.
There are some moderately finicky details that must be observed to
get this right and keep it going over time, but this configuration has
made it possible for my group to maintain nearly 100% file accessibility
through both scheduled and unscheduled server downtime.
> I should add that we are lagging/lingering on 3.4a, which I fully realise
> is far from being the latest-and-greatest release.
I, at least, also found limited incentive to upgrade to 3.5 since it's
scheduled to become unsupported at the same time as 3.4a (Jan 31, 2001).
We found AFS 3.5 to be an unsually painful upgrade because Sun's support
for 2.6 was sufficiently poor that we held our servers at 2.5.1 until we
were ready for Solaris 7 (aka 2.7, sun4x_57); as a result we have to
upgrade OS and AFS at the same time.
Regardless, we finally made the switch to 3.5 database servers
this summer so that we can migrate fileservers to 3.6 between now and the
end of January without relying on servers that differ by more than a
single minor release.
I'm still concerned that the 3.6 support period [aka "Program Services"]
is only through June 30, 2001, but we're operating on the presumption that
this will be extended.
Charles Ball
Information Technology
Boston University
